Command line reference

For operators and contributors. Reference.

Generated page

This page is written by site/scripts/generate.mjs from the argument parser in backend/src/cso/cli.py. Do not edit it by hand; change the code and rebuild.

The cso command is installed with the backend. Run it from the backend folder with uv run cso <command>. Add --help to any command to see the same text in your terminal.

Commands

Command	What it does
cso calibrate	refit probe cut points from reviewers’ decisions (dry run by default)
cso api	run the HTTP API
cso worker	run the job worker
cso record	record an anonymized replay fixture from the live OSM API (GET only)
cso db	database maintenance
cso seed	load fixture cases and snapshots for replay mode
cso auth	reviewer access
cso eval	evaluation harness and Day-1 spike

Arguments

cso calibrate

refit probe cut points from reviewers’ decisions (dry run by default)

Argument	What it does	Details
--pack	pack id (default: the active pack)
--dry-run	report only (the default)	default False
--apply	store and activate the result	default False

cso api

run the HTTP API

Argument	What it does	Details
--host		default 127.0.0.1
--port		default 8000

cso worker

run the job worker

Argument	What it does	Details
--once	process due jobs, then exit	default False
--fixture	replay demo: enqueue this fixture (e.g. 02-good-faith-mistake) and run it
--concurrency	jobs in flight at once	default 2

cso record

record an anonymized replay fixture from the live OSM API (GET only)

Argument	What it does	Details
changeset_id
--out	new directory, e.g. fixtures/cases/09-name	required

cso db upgrade

apply migrations (default: head)

Argument	What it does	Details
--revision		default head

cso seed

load fixture cases and snapshots for replay mode

Argument	What it does	Details
--fixtures	directory containing cases/ (default: the recorded demo set)	default fixtures/demo

cso auth create-invite

create an invite code (printed once, stored only as a hash)

Argument	What it does	Details
--role		one of reviewer, admin; default reviewer
--label	who the code is for (not secret)	default
--max-uses	logins allowed; default no limit
--ttl-hours	hours until the code expires; default never

cso eval candidates

collect up to 200 candidate changesets (read-only GETs, polite)

Argument	What it does	Details
--limit		default 200
--max-minutes		default 25.0

cso eval split

freeze the dev/held-out split (written once)

Argument	What it does	Details
--dev		default 60
--heldout		default 120
--dry-run	show sizes and hash without freezing anything	default False

cso eval label

label items from rendered evidence (no model output)

Argument	What it does	Details
--set		one of dev, heldout; default dev
--labeller		required
--session		required
--limit
--legacy	schema version 1: nature and response only, no probes	default False

cso eval adjudicate

merge two labelling sessions

Argument	What it does	Details
--set		one of dev, heldout; default dev
--a		required
--b		required
--adjudicator		required

cso eval spike

Day-1 go/no-go on the 60 dev items

Argument	What it does	Details
--contract		one of panel, permutation; default panel

cso eval panel-diagnostic

run the fact panel over items without labels (not a verdict)

Argument	What it does	Details
--set		one of all, dev, heldout; default all
--profile		one of compact, full; default compact
--include-heldout	required for –set heldout, and for –set all once the split is frozen: sends the held-out ids through the Decider (no labels read) and says so in the report	default False
--tag	suffix for the output names, e.g. -v2	default
--compare-with	append a before/after section for new_account_first_edits against this run

cso eval ablate-model

ask a second Decider server the same panel questions and compare (dev set only)

Argument	What it does	Details
--set		one of dev; default dev
--decider-url	second server, on this machine	required
--label	short model name used in file names	required
--baseline-label	name for the default server’s run	default 2B
--baseline-json	reuse a saved run (a results .json) instead of asking the default server again

cso eval backfill-counts

patch cached snapshots with the edit-time changeset count

No arguments.

cso eval run

run all systems; heldout needs –released

Argument	What it does	Details
--set		one of dev, heldout, fixtures; default dev
--released	required for heldout, after the freeze commit; use for the one final run	default False
--contract		one of panel, permutation; default panel
--confirm-freeze	heldout only: the commit hash that froze evaluation/splits/ (verified against git)
--ablations	heldout only: also run the without-cues and three paraphrase ablations	default False
--authorize-bedrock-usd	heldout only: dollar ceiling that allows the Haiku baselines (also needs CSO_MODE=live and CSO_LIVE_BEDROCK=1); without it they are reported as not run
--hobson-url	heldout only: a second Decider server running hobson-v21 for that ablation
--expect-config-version	heldout only: refuse unless the community config version starts with this

cso eval report

render one markdown report from the latest results files (runs nothing)

Argument	What it does	Details
--latest		default False; required
--out	write here instead of stdout